1 #load MNIST data 2 import tensorflow.examples.tutorials.mnist.input_data as input_data 3 mnist = input_data.read_data_sets("MNIST_data/",one_hot=True) 4 5 #start tensorflow interactiveSession 6 import tensorflow as tf 7 sess = tf.InteractiveSession() 8 9 #weight initilization 10 def weight_variable(shape): 11 initial = tf.truncated_normal(shape, stddev=0.1) 12 return tf.Variable(initial) 13 14 def bias_variable(shape): 15 initial = tf.constant(0.1, shape= shape) 16 return tf.Variable(initial) 17 18 #convolution 19 def conv2d(x, W): 20 return tf.nn.conv2d(x, W, strides=[1,1,1,1], padding='SAME') 21 22 #pooling 23 def max_pool_2x2(x): 24 return tf.nn.max_pool(x, ksize=[1,2,2,1],strides=[1,2,2,1], padding='SAME') 25 26 #Create the model 27 #placeholder 28 x = tf.placeholder("float",[None, 784]) 29 y_ = tf.placeholder("float", [None, 10]) 30 31 #variable 32 W = tf.Variable(tf.zeros([784,10])) 33 b = tf.Variable(tf.zeros([10])) 34 35 y = tf.nn.softmax(tf.matmul(x,W) +b) 36 37 #first convolutional layer 38 w_conv1 = weight_variable([5,5,1,32]) 39 b_conv1 = bias_variable([32]) 40 41 x_image = tf.reshape(x,[-1,28,28,1]) 42 43 h_conv1 =tf.nn.relu(conv2d(x_image,w_conv1) + b_conv1) 44 h_pool1 =max_pool_2x2(h_conv1) 45 46 #second convolutional layer 47 w_conv2 = weight_variable([5,5,32,64]) 48 b_conv2 = bias_variable([64]) 49 50 h_conv2 =tf.nn.relu(conv2d(h_pool1, w_conv2) + b_conv2) 51 h_pool2 =max_pool_2x2(h_conv2) 52 53 #densely connected layer 54 w_fc1 = weight_variable([7*7*64, 1024]) 55 b_fc1 = bias_variable([1024]) 56 57 h_pool2_flat = tf.reshape(h_pool2, [-1,7*7*64]) 58 h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, w_fc1) + b_fc1) 59 60 #dropout 61 keep_prob = tf.placeholder("float") 62 h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob) 63 64 #readout layer 65 w_fc2 = weight_variable([1024,10]) 66 b_fc2 = bias_variable([10]) 67 68 y_conv = tf.nn.softmax(tf.matmul(h_fc1_drop,w_fc2) + b_fc2) 69 70 #train and evaluate the model 71 cross_entropy = -tf.reduce_sum(y_*tf.log(y_conv)) 72 #train_step = tf.train.AdagradOptimizer(1e-4).minimize(cross_entropy) 73 train_step = tf.train.GradientDescentOptimizer(1e-4).minimize(cross_entropy) 74 correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1)) 75 accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float")) 76 sess.run(tf.initialize_all_variables()) 77 for i in range(5000): 78 batch = mnist.train.next_batch(50) 79 if i%100 ==0: 80 train_accuracy = accuracy.eval(feed_dict={x:batch[0],y_:batch[1], keep_prob:1.0}) 81 print "step %d, train accuracy %g " %(i,train_accuracy) 82 train_step.run(feed_dict={x:batch[0],y_:batch[1], keep_prob:0.5}) 83 84 print "test accuracy %g" % accuracy.eval(fedd_dict={x:mnist.test.images, y_:mnist.test.labels, keep_prob:1.0})
同样是极客学院的课程,其实也是翻译的国外的robot-ai博客上的内容,但是这个博客,现在打不开了,可能是墙的问题?没有太深究。
按照作者的说法,是采用自适应下降的方式,在train阶段能达到99%的正确率,但是,我的结果只有93%左右,修改梯度步长到1e-4也只有94% 。因此尝试换用原来的梯度下降方式,反而能获得97.61%的正确率,在训练中还达到过98%,这个问题比较无奈,修改步长的结果提升也并不明显。有人在评论中说在不同的平台上测试的值不同,比如在纯CPU环境,和我的结果比较相似。在K20环境中能达到99%,这个问题留待以后探索。代码参考至:文章链接: http://blog.csdn.net/yhl_leo/article/details/50624471
本站文章如无特殊说明,均为本站原创,如若转载,请注明出处:TensorFlow入门——MNIST深入 - Python技术站